Marketing Agent proposes and executes campaigns using only the data within your Workspace, and it never uses your data to retrain AI models. The sections below detail the scope of data the agent works with and the safeguards Lumo has in place.
Data the agent references
Marketing Agent references only the data within the same Workspace when it makes proposals and performs analysis. It never references data from other Workspaces.
| Type of data | What is referenced |
|---|---|
| Broadcast data | The content of past Message Broadcasts, open rates, click rates, and send dates and times |
| Friends data | The number of friends, the makeup of your Segments, and the distribution of tags |
| Workflow data | The settings and execution status of active Workflows |
| Auto Send Messages | The content and delivery status of configured Auto Send Messages |
| External service data | Shopify products and purchase trends, HubSpot segments (lists), and the like (only when the relevant service is integrated) |
Input to the agent is limited to the data in the target Workspace. The results of processing are also stored only within that Workspace.
Data the agent does not reference
The following data is not used in the agent's proposals or analysis. Privacy-related information is excluded from the agent's input by design.
| Data not referenced | Description |
|---|---|
| Personally identifiable information | Email addresses, phone numbers, addresses, and similar details. These values are not used in proposals or analysis. When the agent summarizes a friend's conversation, email addresses and phone numbers are masked |
| One-to-one chat content | Not used in proposals or analysis. Only when you explicitly ask, such as "Tell me about this friend" or "Analyze the trends in inquiries", does the agent read the recent content of the relevant conversation in order to summarize it |
| Workflow content and Klaviyo profiles | For Workflows, only the settings and execution status are referenced; for Klaviyo, only the integration status |
| Data from other Workspaces | Data is separated per Workspace, and Workspaces cannot access one another |
Data the Customer Agent handles
To answer an incoming message from a friend, the Customer Agent passes the message text, the history of the same conversation, the tail end of the previous session, and Knowledge Base search results to the AI. It does not pass the friend's display name, tags, fields, Shopify, HubSpot, or Klaviyo customer data, images, or stickers.
What the agent learns
The agent learns your brand's tendencies and preferences from your activity in the Workspace, improving the accuracy of its proposals. What it learns is used only within that Workspace.
| What is learned | Description |
|---|---|
| Brand tone | It learns a writing style that fits your brand from past broadcast text and your edits |
| Send timing | It learns which time slots get the best response from open rate data |
| Segment performance | It learns which Segments respond to which themes |
| Direction of preferences | It learns your preferred direction for proposals from your edits and feedback |
What it learns is separated per Workspace. Learning in one Workspace never affects proposals in another Workspace.
Your data is not used to retrain AI models
Your data is never used to train or tune the AI models themselves. Lumo only learns tendencies that optimize proposals within the target Workspace, and your data is never used beyond that scope.
Data security
The AI agent features include safeguards for handling your data securely.
| Item | Details |
|---|---|
| AI platform | Runs on Google Cloud's Vertex AI |
| Processing region | Your data is processed in the Tokyo region. The exceptions are PDF document OCR (US region), described below, and image generation (Google Cloud's global locations) |
| Data separation | Fully separated per Workspace |
| External transmission | By default, your data is never sent to AI services outside Google Cloud. When web search is enabled, only search queries are sent to an external search service (no customer data). If you set up an integration with an external AI agent for the Customer Agent, conversations are passed to that endpoint |
| Retraining | Your data is never used to retrain AI models |
| Quality evaluation | To maintain quality, agent responses are scored automatically within Lumo and may be reviewed by the operations team. They are not used for training |
| Activity records | Conversation excerpts included in the agent's activity records are erased within 90 days. If you delete your Workspace, they are erased at the time of physical deletion following the 30-business-day grace period (only metrics such as counts remain) |
| Action approval | High-impact actions, such as sending a Message Broadcast or activating a Workflow, require approval |
If you add PDF documents to the Knowledge Base, text extraction (OCR) for those files runs in Google Cloud's US region. This covers only the content of the PDF files you upload. With the AI that Lumo uses by default, friend data and chat history are never processed outside Japan. If you set up an integration with an external AI agent for the Customer Agent, where conversations are sent and where they are processed follow the endpoint you configured. For where files such as images and videos received on LINE are stored, see "About Lumo's Security".
Your conversation history with the agent is also managed separately per Workspace.
High-impact actions run only after you approve them
When the agent performs a high-impact action, it always asks for your approval. The action is not carried out until you review the details and choose "Execute". Scheduling and sending broadcasts and activating Workflows always require approval regardless of the Trust Score. The agent has no mechanism to carry out these actions itself; its role ends at preparing the draft or the scheduled content. How much the other actions are delegated varies with the Trust Score.
When you disable AI features for your Workspace, all AI data processing stops, including agent responses and sentiment analysis.
The main actions that require approval are as follows.
- Sending or scheduling a Message Broadcast
- Activating a Workflow
- Creating or editing a Segment
In the approval dialog, you can review the details of the action and then choose "Execute" or "Not now". A Message Broadcast is not sent until you preview the content and approve it.
Adjusting the scope of automation with the Trust Score
The more of the agent's proposals you approve, the higher the Trust Score, which measures the level of autonomy for each action, rises (as a guide: about 5 approvals for Notification Only, 20 for Automatic). Actions where trust has grown skip the per-run approval and proceed automatically. There's currently no screen for changing the Trust Score manually.
Resetting what the agent has learned or the Trust Score
You can delete your own conversation history and learned memory in bulk from Personal Settings. Use this when a team member or your brand's direction changes and you want to start learning over.
Clearing what has been learned
"Clear conversation history and memory" under "AI Agent" in "Personal Settings" deletes the conversation history and learned memory between you and the agent in bulk. There's no list view or per-item deletion, and it doesn't affect other Members' history.
What you clear cannot be restored. Before running it, make sure you really want to erase the data.
Resetting the Trust Score
Resetting the Trust Score isn't currently available from the user screens. If you want an operation you've delegated back on a confirmation basis, contact support.